{"paper":{"title":"Learning Reasoning Rewards from Expert Demonstrations with Inverse Reinforcement Learning","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Inverse reinforcement learning extracts reusable process rewards from expert reasoning traces that improve language model training and inference beyond imitation.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Claudio Fanconi, Mihaela van der Schaar, Nicol\\'as Astorga","submitted_at":"2025-10-02T09:55:26Z","abstract_excerpt":"Teaching large language models (LLMs) to reason during post-training typically relies on reinforcement learning with explicit outcome- or process-based reward functions. However, in many real-world settings, obtaining or defining such reward functions is difficult, especially for complex tasks, making learning from expert demonstrations an attractive alternative. The dominant approach, supervised fine-tuning (SFT), trains models to imitate expert reasoning traces directly, but suffers from the general limitations of off-policy learning: performance can be fragile to inference-time deviations f"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Through experiments on GSM8K, MMLU-Pro and MedReason we show that the reasoning reward function learned with R-AIRL can be effectively used throughout the training and inference pipeline: (1) to provide a training signal for post-training, outperforming SFT in most of the considered settings, (2) for inference-time reranking, improving pass@1 by up to 17.4 points, and (3) for process-level evaluation, localising reasoning failures with up to 86.1% accuracy.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the adversarial inverse RL procedure can recover a generalizable process-level reward from expert demonstrations that truly reflects reasoning quality and transfers to states not explicitly present in the training traces.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"R-AIRL learns a reasoning reward function from expert demonstrations using adversarial inverse RL and shows it outperforms SFT for training while enabling reranking and failure localization on GSM8K, MMLU-Pro, and MedReason.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Inverse reinforcement learning extracts reusable process rewards from expert reasoning traces that improve language model training and inference beyond imitation.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"c9af456146c53a0d44efee218276fc18ea3f53d92fe3ec58b97efe1139437c0f"},"source":{"id":"2510.01857","kind":"arxiv","version":5},"verdict":{"id":"51d263a0-f13d-43bd-be54-1abd5bc77b61","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-18T11:08:44.647830Z","strongest_claim":"Through experiments on GSM8K, MMLU-Pro and MedReason we show that the reasoning reward function learned with R-AIRL can be effectively used throughout the training and inference pipeline: (1) to provide a training signal for post-training, outperforming SFT in most of the considered settings, (2) for inference-time reranking, improving pass@1 by up to 17.4 points, and (3) for process-level evaluation, localising reasoning failures with up to 86.1% accuracy.","one_line_summary":"R-AIRL learns a reasoning reward function from expert demonstrations using adversarial inverse RL and shows it outperforms SFT for training while enabling reranking and failure localization on GSM8K, MMLU-Pro, and MedReason.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the adversarial inverse RL procedure can recover a generalizable process-level reward from expert demonstrations that truly reflects reasoning quality and transfers to states not explicitly present in the training traces.","pith_extraction_headline":"Inverse reinforcement learning extracts reusable process rewards from expert reasoning traces that improve language model training and inference beyond imitation."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2510.01857/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":1,"snapshot_sha256":"be01e8ccc0eb607e8627d867ccb4d8e809ece5b944f26d198b08080121140881"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}